Re: Preview: portable dumper

[Daniel Colascione]

On 02/14/2018 01:03 PM, Philipp Stephani wrote:
> Eli Zaretskii <address@hidden <mailto:address@hidden>> schrieb am Di., 13. 
> Feb. 2018 um 17:39 Uhr:
>
>      > From: Daniel Colascione <address@hidden
>     <mailto:address@hidden>>
>      > Date: Mon, 12 Feb 2018 12:18:36 -0800
>      >
>      > I've pushed the portable dumper to the pdumper Savannah branch. It
>      > should support a fully PIC Emacs.
>
>     Thanks.  I'd urge people to try this branch and report any issues they
>     see.
>
>
> Got a crash when building on Debian testing. I've attached the output of 
> 'bt full'.

Thanks. There were two bugs, both relating to conservative GC:

1) To make conservative GC work, we record in the dump the start 
position and Lisp type of each object in the dump. The problem was that 
we were recording object-start records for objects in the discardable 
region. With the right stack garbage as input, conservative GC would 
find one of these object-start records, use it as evidence that it was 
looking at a real object, then call mark_object, which died because it 
knows that objects in the discardable region shouldn't be live. Fixed by 
just not emitting these discardable-region object-start records anymore.

2) When evaluating a Lisp_Object on the stack in mark_maybe_object, 
occasionally, we can find a word with a valid pointer component but the 
wrong type tags. Now we check that the Lisp_Object's tagged type matches 
the actual type of the object in the dump.

I don't know why it was so hard to repro these on my machine. I had to 
bootstrap a few times before Emacs began crashing.